Segmenting thin, elongated structures such as blood vessels or road networks is notoriously difficult because a single local error can sever an otherwise continuous path. The new arXiv paper introduces TRACE, a topology-aware reconnection method designed to address this fragility. Rather than treating connectivity as a post-processing step, TRACE integrates it into the learning objective.
The authors note that existing skeleton-based objectives typically supervise a fixed centerline, which can be mismatched with the model's evolving predictions. TRACE instead enforces connectivity adaptively, using the prediction itself to guide where reconnection is needed. This shift from static supervision to dynamic enforcement is the core contribution.
While the abstract is brief, the motivation is clear: preserving topology in thin structures requires objectives that respond to the model's current output, not just a precomputed ground-truth skeleton. The paper is available on arXiv under identifier 2610.22210.